AI PM @Memoket_AI Building AI memory. Exploring what's next in AI. Product & tech deep dives. Follow for the product, stay for the ideas and conversations.
Joined July 2026
- Tweets266
- Following209
- Followers56
- Likes549
Your recordings in Claude Code
Memoket users: connect your recordings to Claude Code:
claude mcp add --transport http memoket mcp.memoket.ai/mcp
Run /mcp, select memoket, sign in.
Try: “Find recordings where we discussed the launch date.”
Read-only. Connectors is free; your Claude plan is separate.
Optional documentation self-reply: Connectors setup
1
3
72
Binyu retweeted
I wrote a little about the importance of preparing yourself for any opportunity. Don't let the train pass right in front of your eyes.
Binyu retweeted
最近有个感受：很多时候我们的想法可能没那么重要。因为想法一天一变，真正一天就可以有很多想法。真正重要的还是行动，是你在一件事上投入了多少时间，自己亲身经历了什么。
每个人的认知，其实都在受到自己曾经做过的事情的影响。比起反复思考目标、计划未来，为还没发生的事情焦虑，对我们影响更大的，可能是做了什么事，见了什么人，遇到了什么问题，又得到了什么反馈。
人是有神经可塑性的，我们的大脑会随着经历和学习，持续调整神经连接，甚至发生结构上的变化。虽然思考本身也能改变大脑，但真实的行动往往会带来更多意想不到的刺激和反馈。
大脑的改变需要一些“摩擦”，单纯的想法本身很难构成有力的摩擦。与其总想着未来要成为什么样的人，人在AI时代是不是没有机会了，不如行动起来聚焦在做事上，接触不同的人，创造一些新的经历。具身的体验和拿反馈，是AI现阶段还无法替代的。
有句老话“听过了很多道理，依然过不好这一生”，可能就是因为很多事情只停留在想法阶段，没有行动的摩擦，也没有现实的反馈吧。
AI is changing how we discover products and information. Instead of clicking through search results, more people are getting answers directly from AI.
That raises an interesting question: If people stop clicking, how will they discover your brand?
We recently shared our thoughts on this shift from traditional SEO to GEO (Generative Engine Optimization), and why we're focusing on real user stories, original insights, and content that AI can actually reference.
📷 Read our latest blog: memoket.ai/blogs/news/google…
📷 Curious what you think: Have you started using ChatGPT or other AI tools instead of Google to discover new products?
📝 We read every review—including the cons.
Over the past few months, journalists have put Memoket Gem to the test. They raised fair questions about subscription pricing, audio quality in larger rooms, charging, and security.
Some of those criticisms were right. Here’s what we’re doing about them:
✅ Membership pricing is now public.
🎧 Noise Reduction is live in V1.3.3, improving playback clarity in noisy environments.
🔌 A charger that works with standard USB-C cables is in development.
🔒 SOC 2 Type II audit completion is expected in early November.
We’re also working on better microphones for Gem 2. Noise Reduction helps with background noise, but better capture range needs better hardware.
The post links to the original reviews and explains what’s improved, what’s in progress, and where limitations remain.
👉 Read our response: memoket.ai/blogs/news/we-rea…
💬 What’s one thing you’d most like us to improve? We’d love to hear it.
Your latest conversation has a backstory.
Find the connections with Memoket’s Related Context.
Read Basic Tutorials’ review:
basic-tutorials.com/reviews/…
MemoketはClaudeのコネクターにも対応しています。ClaudeのディレクトリでMemoketを追加し、サインインして許可したあと、チャットの「＋」→「Connectors」でMemoketをオンにすると、録音の検索や文字起こしの参照ができます。読み取り専用です。プランや設定によって使えない場合があります。
1
42
Binyu retweeted
KITE's experimental remember() needs one writer.
It rewrites a single XML file; concurrent writes aren't supported.
Save session_id + date with each conversation. Check fact.sources when recalling it.
API: github.com/memoket/memoket-k…
The exchange example is a useful test: if a user narrows a two-item request to one, does the agent respect the update? Both calls can pass the API. Training on the whole 'successful' run can miss that difference.
Trajectory2Gym: From successful trajectories to grounded decisions
A new framework from Ant Group for training tool-use agents. It constructs composable interactions and supervises only actions justified by visible history. Across five Qwen3 scales, it improves Tau1, Tau2, and BFCL by up to +29.8 pp.
35
Many memory systems lost to plain RAG over the raw history, and the losses traced to what memory did to the evidence: dropped, fragmented, rephrased. So the question isn't how clever the memory layer is, it's whether each memory can still point back to its source.
[1/4] 🤯 Many agent memory systems can’t even beat a naive RAG?
We conducted a comprehensive evaluation and find many underperform the naive RAG, even at higher cost.
To better understand this, we introduce “Memory Tax”, see more in our paper.
arxiv.org/abs/2610.34073
37
エージェントの記憶で私が大事だと思うのは、何を覚えたかより、どこから覚えたかを確かめられることです。Memoketでは、AIの答えに元の会話へのリンクを付けています。間違っていそうなら、元の会話に戻って確認できます。
25
インタビューを録音するか問題。短い記事なら録らなくても書けるのは確かにそうだと思います。録音が効くのは、引用する一言を本人の言い回しのまま確かめたい時。会話を録ってAIに渡すMemoketのPMとしても、要約より元の言い回しに戻れることを大事にしています。
インタビューを録音するかしないか？
僕は必ずする。しないなんてあり得ない……と思ったけど、よく考えたら記事の内容と分量にもよるよね。
特に新聞、雑誌だと1000字以下の短いインタビュー記事も珍しくない。それなら確かに録音しないでも印象に残ったところを書けば終わる。
僕も数年前、雑誌で短いインタビュー記事を書いていた時は、録音しても文字起こしせず、記憶を頼りに原稿を書いていて、それでなんの問題もなかった。
取材のテーマが決まっていて、原稿の文字数が少なければ、録音せずメモで済ませることも十分に可能だろう。
でも、書けるかどうかとは別に、リスクヘッジとして録音するべきだと思う。どんなきっかけで言った言わないのトラブルが発生するかわからないからね（実際にあった）。自分を守る保険として音声データは大切。
#ライター
29
Gemの使い方はシンプルです。録音ボタンを1秒長押しすると録音が始まり、もう一度押すと止まります。録音中は赤いLEDが点灯します。スマホのアプリと同期すると、文字起こしとAI要約をアプリで確認できます。文字起こしは153の言語・地域別バリエーションに対応しています。
51
LINEの通話で録音と文字起こしが使えるようになると、記録を残すこと自体はぐっと身近になります。差が出るのは、後から「誰が何を決めたか」をすぐ探せるかどうかだと思っています。対面の会議でそこを作っているので（MemoketのPMです）、気になるニュースでした。
［ITmedia Mobile］LINEで通話の録音／録画が可能に 音声通話は文字起こしも 今冬からは有料に itmedia.co.jp/mobile/article…
1
66
"Search is a series of decisions" fits conversations too. With recordings, the first decision is often who said it, then when. At Memoket, our connector has a separate tool for that: transcripts by participant, next to search across recordings. All read-only.
▲ Nace AI just released Drex for Search, which brings its Drex decision models to document search, so AI agents get cited pages instead of top-k chunks.
The key idea is that search is a series of decisions. Drex first classifies the question as a quick lookup or a deeper agent search. It then narrows the workspace from folder to file, section and page, down to table and column inside a sheet. Only the kept pages reach the LLM, and the answer cites the file and page.
The numbers, per Nace: 92.05% on BrowseComp-Plus with GPT-5, at 12.61 search calls per question, 16.6x fewer than the only entry ahead. 77.5% R@1 on VisDocAgentBench, 10 points above Claude Opus 5. 94.7% on MADQA with Gemini 3.5 Flash. On OfficeQA Pro, all 9 agent setups Nace tested needed fewer turns, with Claude Code on Opus 5.5 going from 18.5 to 9.1 turns.
Claude Code, Codex, Cursor and OpenCode can call it through the nace-mcp server or the CLI. There are also Python and TypeScript SDKs and a REST API......
Full analysis: marktechpost.com/2026/10/26/…
Product: aicampaign.live/NC-A1-1
Console: aicampaign.live/nc-a1-2
Docs: console.nace.ai/docs
@NaceAI
2
35
Every execution suggestion adopted, no strategy one, yet a single human sentence mid-run beat going alone on every benchmark. So the design question isn't more memory. It's when to ask. An agent that knew its strategy was a default, not a decision, could flag it for review.
An agent that can run a full post-training pipeline almost never reconsiders its initial strategy.
We analyze 1,338 complete training trajectories from PostTrainBench — 7 benchmarks, 4 base models, Claude Code / Codex CLI / OpenCode, 10h on one H100 — and split capability into EXECUTION vs STRATEGY.
1️⃣ Qualified executor, locked strategy. Agents improve over base models on every benchmark; with experience, avg score goes 33.34 → 48.55. ALL execution-level suggestions were adopted, but NONE of the strategy-level suggestions (e.g., switching to RL, adding SFT warmup) were adopted. Only 2.1% of 3,557 adjacent experiment pairs actually switch strategy. Defaults follow the agent, not the task: Claude Code starts full-parameter SFT 71.9% of the time; Codex CLI starts PEFT/LoRA 89.5%.
2️⃣ Experience and compute don't unlock it. Extra experience and memory raise execution but never leave the locked strategy; more reasoning tokens buy finer execution rather than better direction — on AIME 2025, ~14M extra tokens solve just 1 more problem.
3️⃣ A mid-run human guidance wins everywhere. At the same checkpoint, same remaining budget, one sentence switching to the alternative strategy beats autonomous continuation: GSM8K +17.44pp, HumanEval +10.80pp, AIME 2025 +6.67pp.
The missing ability isn't executing a new strategy — it's the trigger to spontaneously restart strategy decisions mid-trajectory.
📄 Paper: arxiv.org/abs/2608.19072
#AIAgent #PostTraining #LLM #RLHF #GRPO #PPO #RSI #AgentBenchmark #PostTrainBench #THUNLP
1
78
AI agents can read files, code and chats. But context also lives outside them: a customer call, a hallway conversation, an idea you said out loud.
At Memoket, we connect recorded conversations to the AI tools you already use, so real-world context can travel with digital context.
2
49
